Time series models struggle with simple patterns despite training
Foundations without Fundamentals: Zero-Shot Blind Spots in Time Series FMs
Machine Learning
Summary
Time series foundation models are designed to forecast data points over time, but this paper finds that popular models often fail to learn very basic patterns like steady trends or simple cycles without extra training. The authors created a set of simple tests called SimpleTimeBench to reveal these weaknesses. Even when fine-tuned on some tasks, these models often lose their ability to handle other basic patterns, showing a gap in their understanding of time logic. These problems also appear in real-world sensor data, where models miss obvious predictor signals, limiting their reliability.
What this means in practice
- •For industrial sensor data teams: Identify when multivariate time series models fail to utilize predictive sensor signals in zero-shot forecasting to improve operational reliability.
- •For financial forecasting teams: Detect limitations in current time series models for capturing simple leading indicators, allowing better model selection or additional training.
Authors
Nafiseh Ghoroghchian, Haipeng Zhang, Shuyi Han, Alex Labach, George Stein
Abstract
Despite the success of Time Series Foundation Models (TSFMs) on broad benchmarks, their ability to internalize basic temporal logic, especially in settings supported by exogenous covariates, remains under-examined. We introduce SimpleTimeBench, a diagnostic univariate and multivariate "unit test" suite for primitives such as monotonic trends, periodic signals and leading indicator covariates, scenarios where near-perfect forecasts should be trivial. Surprisingly, prominent multivariate TSFMs (Chronos-2, Moirai and Toto) frequently produce suboptimal zero-shot forecasts for these inputs. While fine-tuning Chronos-2 improves its behaviour on specific tasks, we show that this adaptation degrades performance on other fundamental patterns rather than enhancing its generalizable foundational capabilities. This reveals a gap between pre-training scale and basic temporal reasoning, suggesting that current TSFMs could potentially lack the inductive biases needed to capture simple predictable functions. We further demonstrate that these failures are not merely synthetic curiosities: they persist in real-world sensor forecasting, where TSFMs consistently underutilize leading indicators available in observed covariates. This inability to capture simple relationships limits the practical utility and reliability of current multivariate models.